Parameter Optimization in VAD Used in Noise Spectral Estimation for Spectral Suppression Method

摘要

In spectral suppression noise canceller, spectral gain is multiplied to the noisy speech spectrum to reduce noise. To calculate spectral gain, noise spectrum estimation is necessary. By using Voice Activity Detector (VAD), noise spectrum is estimated by discriminating the input noisy speech into speech/quasi-speech/non-speech frames. The threshold of noisy speech discrimination is not adaptive to input noise type in the conventional method. In this paper, threshold optimization method is proposed. Based on noise characteristics, optimal parameters for colorless noise and colored noise are proposed. Also, automatic detection of colorless noise and colored noise algorithm is proposed. Simulations are carried out in 5 different noise-changed conditions by using 3 sets of input speech signals.